Convolutional Neural Networks

Project: Write an Algorithm for a Dog Identification App


In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!

Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the Jupyter Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.

In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.

Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.

The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this Jupyter notebook.


Why We're Here

In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).

Sample Dog Output

In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!

The Road Ahead

We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.

  • Step 0: Import Datasets
  • Step 1: Detect Humans
  • Step 2: Detect Dogs
  • Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
  • Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)
  • Step 5: Write your Algorithm
  • Step 6: Test Your Algorithm

Step 0: Import Datasets

Make sure that you've downloaded the required human and dog datasets:

Note: if you are using the Udacity workspace, you DO NOT need to re-download these - they can be found in the /data folder as noted in the cell below.

  • Download the dog dataset. Unzip the folder and place it in this project's home directory, at the location /dog_images.

  • Download the human dataset. Unzip the folder and place it in the home directory, at location /lfw.

Note: If you are using a Windows machine, you are encouraged to use 7zip to extract the folder.

In the code cell below, we save the file paths for both the human (LFW) dataset and dog dataset in the numpy arrays human_files and dog_files.

In [1]:
import numpy as np
from glob import glob

# load filenames for human and dog images
human_files = np.array(glob("/data/lfw/*/*"))
dog_files = np.array(glob("/data/dog_images/*/*/*"))

# print number of images in each dataset
print('There are %d total human images.' % len(human_files))
print('There are %d total dog images.' % len(dog_files))
There are 13233 total human images.
There are 8351 total dog images.

Step 1: Detect Humans

In this section, we use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images.

OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory. In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.

In [2]:
import cv2                
import matplotlib.pyplot as plt                        
%matplotlib inline                               

# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')

# load color (BGR) image
img = cv2.imread(human_files[0])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# find faces in image
faces = face_cascade.detectMultiScale(gray)

# print number of faces detected in the image
print('Number of faces detected:', len(faces))

# get bounding box for each detected face
for (x,y,w,h) in faces:
    # add bounding box to color image
    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
    
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
Number of faces detected: 1

Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.

In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.

Write a Human Face Detector

We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.

In [3]:
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
    img = cv2.imread(img_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray)
    return len(faces) > 0

(IMPLEMENTATION) Assess the Human Face Detector

Question 1: Use the code cell below to test the performance of the face_detector function.

  • What percentage of the first 100 images in human_files have a detected human face?
  • What percentage of the first 100 images in dog_files have a detected human face?

Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.

Answer: (You can print out your results and/or write your percentages in this cell)

In [4]:
from tqdm import tqdm
from functools import reduce

human_files_short = human_files[:100]
dog_files_short = dog_files[:100]

#-#-# Do NOT modify the code above this line. #-#-#
In [7]:
def percent_human(files, detector):
    detected = list(map(detector, files))
    return sum(map(lambda x: 1 if x else 0, detected)) / len(detected)

print("percent of humans detected to be humans: {}".format(percent_human(human_files_short, face_detector)))
print("percent of dogs detected to be humans: {}".format(percent_human(dog_files_short, face_detector)))
percent of humans detected to be humans: 0.98
percent of dogs detected to be humans: 0.17

We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.

In [5]:
import time
from os import listdir

cascades = { cascade:file for [cascade, file] in map(lambda x : [x[24:-4], 'haarcascades/' + x], listdir('haarcascades')) }

def detect_object_cascade(cascade_file, img_path):
    cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
    img = cv2.imread(img_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    objects = cascade.detectMultiScale(gray)
    return len(objects) > 0

def percent_detected(img_files, detector):
    detected = list(map(detector, img_files))
    return sum(map(lambda x: 1 if x else 0, detected)) / len(detected)

def rate_detector(detection_method, detector):
    start = time.time()
    print("rating detection method {}".format(detection_method))
    print("percent of human images detected: {}".format(percent_detected(human_files_short, detector)))
    print("percent of dogs images detected: {}".format(percent_detected(dog_files_short, detector)))
    print("took {} seconds".format(time.time() - start))
    print("=========")

def cascade_detector(cascade_file):
    return lambda x : detect_object_cascade(cascade_file, x)
In [6]:
for cascade_name, cascade_file in cascades.items():
    rate_detector(cascade_name, cascade_detector(cascade_file))
rating detection method alt2
percent of human images detected: 0.98
percent of dogs images detected: 0.17
took 92.338787317276 seconds
=========
rating detection method alt_tree
percent of human images detected: 0.98
percent of dogs images detected: 0.17
took 90.6826696395874 seconds
=========
rating detection method default
percent of human images detected: 0.98
percent of dogs images detected: 0.17
took 90.64785432815552 seconds
=========
rating detection method alt
percent of human images detected: 0.98
percent of dogs images detected: 0.17
took 90.48003697395325 seconds
=========

Conclusion

None of them were any better than any other in detection and the difference in time it took to run was negligble and could have been due to any number of factors


Step 2: Detect Dogs

In this section, we use a pre-trained model to detect dogs in images.

Obtain Pre-trained VGG-16 Model

The code cell below downloads the VGG-16 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories.

In [ ]:
import torch
import torchvision.models as models

# define VGG16 model
VGG16 = models.vgg16(pretrained=True)

# check if CUDA is available
use_cuda = torch.cuda.is_available()

# move model to GPU if CUDA is available
if use_cuda:
    VGG16 = VGG16.cuda()

print(VGG16)
Downloading: "https://download.pytorch.org/models/vgg16-397923af.pth" to /root/.torch/models/vgg16-397923af.pth
100%|██████████| 553433881/553433881 [00:05<00:00, 98896081.71it/s] 

Given an image, this pre-trained VGG-16 model returns a prediction (derived from the 1000 possible categories in ImageNet) for the object that is contained in the image.

(IMPLEMENTATION) Making Predictions with a Pre-trained Model

In the next code cell, you will write a function that accepts a path to an image (such as 'dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg') as input and returns the index corresponding to the ImageNet class that is predicted by the pre-trained VGG-16 model. The output should always be an integer between 0 and 999, inclusive.

Before writing the function, make sure that you take the time to learn how to appropriately pre-process tensors for pre-trained models in the PyTorch documentation.

In [ ]:
from PIL import Image
import torchvision.transforms as transforms

def classify(model, img_path):
    if use_cuda:
        model = model.cuda()
    
    for param in model.parameters():
        param.requires_grad_(False)
    
    model.eval()
    original_image = Image.open(img_path).convert('RGB')
    
    in_transform = transforms.Compose([
                    transforms.Resize((224, 224)),
                    transforms.ToTensor(),
                    transforms.Normalize((0.485, 0.456, 0.406), 
                                         (0.229, 0.224, 0.225))])
    
    image = in_transform(original_image)[:3,:,:].unsqueeze(0)
    image.unsqueeze(0)
    if use_cuda:
        image = image.cuda()
    
    output = model(image)
    index = torch.max(output, 1)[1][0].item()
    
    return index
    
def VGG16_predict(img_path):
    '''
    Use pre-trained VGG-16 model to obtain index corresponding to 
    predicted ImageNet class for image at specified path
    
    Args:
        img_path: path to an image
        
    Returns:
        Index corresponding to VGG-16 model's prediction
    '''
    return classify(VGG16, img_path)

(IMPLEMENTATION) Write a Dog Detector

While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained VGG-16 model, we need only check if the pre-trained model predicts an index between 151 and 268 (inclusive).

Use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).

In [ ]:
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
    return 151 <= VGG16_predict(img_path) <= 268

(IMPLEMENTATION) Assess the Dog Detector

Question 2: Use the code cell below to test the performance of your dog_detector function.

  • What percentage of the images in human_files_short have a detected dog?
  • What percentage of the images in dog_files_short have a detected dog?

Answer:

In [9]:
rate_detector('VGG16 based dog detection', dog_detector)
rating detection method VGG16 based dog detection
percent of human images detected: 0.0
percent of dogs images detected: 1.0
took 8.139695405960083 seconds
=========

We suggest VGG-16 as a potential network to detect dog images in your algorithm, but you are free to explore other pre-trained networks (such as Inception-v3, ResNet-50, etc). Please use the code cell below to test other pre-trained PyTorch models. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.

In [10]:
import torch
import torchvision.models as models

use_cuda = torch.cuda.is_available()

some_models = {
    'resnet18' : models.resnet18(pretrained=True),
    'alexnet' : models.alexnet(pretrained=True),
    'squeezenet' : models.squeezenet1_0(pretrained=True),
    'vgg16' : models.vgg16(pretrained=True),
    'densenet' : models.densenet161(pretrained=True),
    'inception' : models.inception_v3(pretrained=True),
#Apparently some environments are missing a number of models AttributeError: module 'torchvision.models' has no attribute 'googlenet'
#    'googlenet' : models.googlenet(pretrained=True),
#    'shufflenet' : models. shufflenet_v2_x1_0(pretrained=True),
#    'mobilenet' : models.mobilenet_v2(pretrained=True),
#    'resnext50_32x4d' : models.resnext50_32x4d(pretrained=True),
#    'wide_resnet50_2' : models.wide_resnet50_2(pretrained=True),
#    'mnasnet' : models.mnasnet1_0(pretrained=True)
}

def generate_dog_detector(model):
    return lambda x: 151 <= classify(model, x) <= 268
Downloading: "https://download.pytorch.org/models/resnet18-5c106cde.pth" to /root/.torch/models/resnet18-5c106cde.pth
100%|██████████| 46827520/46827520 [00:00<00:00, 74069423.92it/s]
Downloading: "https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth" to /root/.torch/models/alexnet-owt-4df8aa71.pth
100%|██████████| 244418560/244418560 [00:03<00:00, 81037605.64it/s]
/opt/conda/lib/python3.6/site-packages/torchvision-0.2.1-py3.6.egg/torchvision/models/squeezenet.py:94: UserWarning: nn.init.kaiming_uniform is now deprecated in favor of nn.init.kaiming_uniform_.
/opt/conda/lib/python3.6/site-packages/torchvision-0.2.1-py3.6.egg/torchvision/models/squeezenet.py:92: UserWarning: nn.init.normal is now deprecated in favor of nn.init.normal_.
Downloading: "https://download.pytorch.org/models/squeezenet1_0-a815701f.pth" to /root/.torch/models/squeezenet1_0-a815701f.pth
100%|██████████| 5017600/5017600 [00:00<00:00, 13319144.04it/s]
/opt/conda/lib/python3.6/site-packages/torchvision-0.2.1-py3.6.egg/torchvision/models/densenet.py:212: UserWarning: nn.init.kaiming_normal is now deprecated in favor of nn.init.kaiming_normal_.
Downloading: "https://download.pytorch.org/models/densenet161-8d451a50.pth" to /root/.torch/models/densenet161-8d451a50.pth
100%|██████████| 115730790/115730790 [00:01<00:00, 81870790.45it/s]
Downloading: "https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth" to /root/.torch/models/inception_v3_google-1a9a5a14.pth
100%|██████████| 108857766/108857766 [00:01<00:00, 68838270.38it/s]
In [11]:
for model_name, model in some_models.items():
    rate_detector("{} based dog detection".format(model_name), generate_dog_detector(model))
rating detection method resnet18 based dog detection
percent of human images detected: 0.01
percent of dogs images detected: 1.0
took 3.8545022010803223 seconds
=========
rating detection method alexnet based dog detection
percent of human images detected: 0.01
percent of dogs images detected: 0.99
took 2.891031265258789 seconds
=========
rating detection method squeezenet based dog detection
percent of human images detected: 0.03
percent of dogs images detected: 1.0
took 3.1622188091278076 seconds
=========
rating detection method vgg16 based dog detection
percent of human images detected: 0.0
percent of dogs images detected: 1.0
took 7.586473226547241 seconds
=========
rating detection method densenet based dog detection
percent of human images detected: 0.0
percent of dogs images detected: 1.0
took 14.627135992050171 seconds
=========
rating detection method inception based dog detection
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-11-2fe9996aa45c> in <module>()
      1 for model_name, model in some_models.items():
----> 2     rate_detector("{} based dog detection".format(model_name), generate_dog_detector(model))

<ipython-input-5-aee2d44d04bd> in rate_detector(detection_method, detector)
     18     start = time.time()
     19     print("rating detection method {}".format(detection_method))
---> 20     print("percent of human images detected: {}".format(percent_detected(human_files_short, detector)))
     21     print("percent of dogs images detected: {}".format(percent_detected(dog_files_short, detector)))
     22     print("took {} seconds".format(time.time() - start))

<ipython-input-5-aee2d44d04bd> in percent_detected(img_files, detector)
     12 
     13 def percent_detected(img_files, detector):
---> 14     detected = list(map(detector, img_files))
     15     return sum(map(lambda x: 1 if x else 0, detected)) / len(detected)
     16 

<ipython-input-10-5eec0c5bc086> in <lambda>(x)
     21 
     22 def generate_dog_detector(model):
---> 23     return lambda x: 151 <= classify(model, x) <= 268

<ipython-input-7-000e6f41d7c7> in classify(model, img_path)
     23         image = image.cuda()
     24 
---> 25     output = model(image)
     26     index = torch.max(output, 1)[1][0].item()
     27 

/opt/conda/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    489             result = self._slow_forward(*input, **kwargs)
    490         else:
--> 491             result = self.forward(*input, **kwargs)
    492         for hook in self._forward_hooks.values():
    493             hook_result = hook(self, input, result)

/opt/conda/lib/python3.6/site-packages/torchvision-0.2.1-py3.6.egg/torchvision/models/inception.py in forward(self, x)
    115         x = self.Mixed_7c(x)
    116         # 8 x 8 x 2048
--> 117         x = F.avg_pool2d(x, kernel_size=8)
    118         # 1 x 1 x 2048
    119         x = F.dropout(x, training=self.training)

RuntimeError: Given input size: (2048x5x5). Calculated output size: (2048x0x0). Output size is too small at /opt/conda/conda-bld/pytorch_1524584710464/work/aten/src/THCUNN/generic/SpatialAveragePooling.cu:63

Conclusion

vgg16 and densenet had the same accuracy, but vgg16 was approximately twice as fast. Looks like inception wants a larger base image. From torchvision docs. I don't think it's worth further investigation at this time. vgg16 seems reasonable.

Important: In contrast to the other models the inception_v3 expects tensors with a size of N x 3 x 299 x 299, so ensure your images are sized accordingly.


Step 3: Create a CNN to Classify Dog Breeds (from Scratch)

Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 10%. In Step 4 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.

We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have trouble distinguishing between a Brittany and a Welsh Springer Spaniel.

Brittany Welsh Springer Spaniel

It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).

Curly-Coated Retriever American Water Spaniel

Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.

Yellow Labrador Chocolate Labrador Black Labrador

We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.

Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!

(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset

Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dog_images/train, dog_images/valid, and dog_images/test, respectively). You may find this documentation on custom datasets to be a useful resource. If you are interested in augmenting your training and/or validation data, check out the wide variety of transforms!

In [9]:
import os
from torchvision import datasets

### TODO: Write data loaders for training, validation, and test sets
## Specify appropriate transforms, and batch_sizes

base_dir = '/data/dog_images'

train_transform = transforms.Compose([transforms.Resize(300),
                                      transforms.RandomResizedCrop(256),
                                      transforms.RandomRotation(30),
                                      transforms.RandomHorizontalFlip(),
                                      transforms.ToTensor(),
                                      transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
                                     ])

test_transform = transforms.Compose([transforms.Resize(300),
                                     transforms.CenterCrop(256),
                                     transforms.ToTensor(),
                                     transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
                                     ])

train_data = datasets.ImageFolder(base_dir + '/train', transform=train_transform)
valiadation_data = datasets.ImageFolder(base_dir + '/valid', transform=test_transform)
test_data = datasets.ImageFolder(base_dir + '/test', transform=test_transform)

train_loader = torch.utils.data.DataLoader(train_data, batch_size=10, shuffle=True)
validation_loader = torch.utils.data.DataLoader(valiadation_data, batch_size=10, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=10, shuffle=True)

loaders_scratch = {
    'train': train_loader,
    'valid': validation_loader,
    'test': test_loader
}

Question 3: Describe your chosen procedure for preprocessing the data.

  • How does your code resize the images (by cropping, stretching, etc)? What size did you pick for the input tensor, and why?
  • Did you decide to augment the dataset? If so, how (through translations, flips, rotations, etc)? If not, why not?

Answer:

I decided to pick a larger image size of 512 x 512, because it looks like we don't have that many files, and so I'm favoring larger images over training speed in hopes of getting richer training. Also, powers of 2 should allow for a good number of transforms. This did not work out as I kept running out of memory. I'm resizing then center cropping for test data. Resizing first should mean that our crop is a significant portion of the image regardless of original size. Center cropping should allow us to throw away some background noise.

For training, I decided to use RandomResizeCrop because it feels like that should do mostly the same thing but expand the dataset through some randomness. I decided to augment the dataset for training in order to extend the training material at least in part because it looks like we don't have that many files. I chose rotation of up to 30 degrees as I expect most images to be relatively vertical, but to have dog heads cocked occasionally. I chose to have horizontal flips since dogs are mostly symmetric, and it should thus effectively double the input data.

I went back and forth on this stuff a bit more as I expanded my research, but ended up mostly back where I started from.

(IMPLEMENTATION) Model Architecture

Create a CNN to classify dog breed. Use the template in the code cell below.

In [10]:
import torch.nn as nn
import torch.nn.functional as F

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        ## Define layers of a CNN
        conv_depths = [[3, 16], [16, 32], [32, 64], [64, 128], [128, 256]]

        self.conv_layers = nn.ModuleList(
            [nn.Conv2d(in_layers, out_layers, 3, padding = 1) for in_layers, out_layers in conv_depths])
        self.batch_norm_layers = nn.ModuleList(
            [nn.BatchNorm2d(out_layers) for in_layers, out_layers in conv_depths])
        self.pool = nn.MaxPool2d(2,2)
        input_xy = 256
        num_pools = len(conv_depths)
        xy_redux = 2**num_pools
        output_xy = int(input_xy / xy_redux)
        conv_depth = conv_depths[-1][1]
        self.global_avg_pool = nn.AvgPool2d(output_xy)
        # without the global average pooling layer the next line is true
        # self.linear_input = conv_depth * output_xy * output_xy
        self.linear_input = conv_depth
        linear_dims = [[self.linear_input, 4096],
                      [4096, 2048],
                      [2048, 1024]]
        
        self.middle_linear_layers = nn.ModuleList(
            [nn.Linear(in_layers, out_layers) for in_layers, out_layers in linear_dims])
        
        self.last_layer = nn.Linear(linear_dims[-1][1], 133)
        self.dropout = nn.Dropout(0.2)
    
    def forward(self, x):
        for i, conv_layer in enumerate(self.conv_layers):
            x = self.batch_norm_layers[i](self.pool(F.relu(conv_layer(x))))
        x = self.global_avg_pool(x)
        x = x.view(-1, self.linear_input)
        self.dropout(x)
        for layer in self.middle_linear_layers:
            x = self.dropout(F.relu(layer(x)))
        return self.last_layer(x)
        

#-#-# You do NOT have to modify the code below this line. #-#-#

# instantiate the CNN
model_scratch = Net()

# move tensors to GPU if CUDA is available
if use_cuda:
    model_scratch.cuda()

print(model_scratch)
Net(
  (conv_layers): ModuleList(
    (0): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (4): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  )
  (batch_norm_layers): ModuleList(
    (0): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (4): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
  (pool): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  (global_avg_pool): AvgPool2d(kernel_size=8, stride=8, padding=0)
  (middle_linear_layers): ModuleList(
    (0): Linear(in_features=256, out_features=4096, bias=True)
    (1): Linear(in_features=4096, out_features=2048, bias=True)
    (2): Linear(in_features=2048, out_features=1024, bias=True)
  )
  (last_layer): Linear(in_features=1024, out_features=133, bias=True)
  (dropout): Dropout(p=0.2)
)

Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step.

Answer:

First, I wanted to be able to change the sizes and number of my layers without having to recalculate everything so I made it based on some configuration.

Next, I was inspired by the vgg16. It had a bit of a larger problem space than just dog breed classification but it seemed to do a pretty good job on dog breed classification. I decided to mimic the idea of multiple convolutional layers per max pooling layer and the 3 linear layers at the end. Since we have a much smaller data set, I figure we have a smaller risk of overfitting and so chose a smaller dropout.

I found a number of hardware limitations. Tried to make the convolutional layers too large and it took too long to just create the net. Then I ended up running out of RAM, and started scaling things back. I also seemed to get a truncated file issue so I followed https://github.com/keras-team/keras/issues/5475 .

My scaled back solution wasn't working. I went through various articles. I found a few implementations of a dog breed classifier in Keras. This one looked useful https://raw.githubusercontent.com/poojasriravichandran/Dog_App/master/dog_app.ipynb . I found it was using this global pooling layer. That lead me to a number of articles, https://keras.io/layers/pooling/ https://discuss.pytorch.org/t/global-average-pooling-in-pytorch/6721 https://alexisbcook.github.io/2017/global-average-pooling-layers-for-object-localization/ <-- by one of our instructors. But the global averaging didn't crack the problem.

A bunch more research and banging my head against the wall, I found https://medium.com/@uijaz59/dog-breed-classification-using-pytorch-207cf27c2031 . They were doing batch normalization. I read up on that https://towardsdatascience.com/batch-normalization-in-neural-networks-1ac91516821c and added it to my network. BAM! Magic bullet to get from 1% accuracy to 47%.

Other things I tried: various batch sizes, various numbers and sizes of convolutional and dense layers, a couple different optimizers.

(IMPLEMENTATION) Specify Loss Function and Optimizer

Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_scratch, and the optimizer as optimizer_scratch below.

In [11]:
import torch.optim as optim

### TODO: select loss function
criterion_scratch = nn.CrossEntropyLoss()

### TODO: select optimizer
optimizer_scratch = optim.Adam(model_scratch.parameters(), lr = 0.001)

(IMPLEMENTATION) Train and Validate the Model

Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_scratch.pt'.

In [12]:
import time

def train(n_epochs, loaders, model, optimizer, criterion, use_cuda, save_path):
    """returns trained model"""
    # initialize tracker for minimum validation loss
    valid_loss_min = np.Inf #0.652
    
    for epoch in range(1, n_epochs+1):
        # initialize variables to monitor training and validation loss
        train_loss = 0.0
        valid_loss = 0.0
        
        ###################
        # train the model #
        ###################
        model.train()
        start = time.time()
        for batch_idx, (data, target) in enumerate(train_loader):
            # move to GPU
            if use_cuda:
                data, target = data.cuda(), target.cuda()
            ## find the loss and update the model parameters accordingly
            ## record the average training loss, using something like
            ## train_loss = train_loss + ((1 / (batch_idx + 1)) * (loss.data - train_loss))
            optimizer.zero_grad()
            output = model(data)
            loss = criterion(output, target)
            loss.backward()
            optimizer.step()
            train_loss += (loss.data - train_loss) / (batch_idx + 1)
            #print('did one')
            
        ######################    
        # validate the model #
        ######################
        model.eval()
        correct = 0.0
        total = 0.0
        with torch.no_grad():
            for batch_idx, (data, target) in enumerate(validation_loader):
                # move to GPU
                if use_cuda:
                    data, target = data.cuda(), target.cuda()
                ## update the average validation loss
                output = model(data)
                loss = criterion(output, target)
                valid_loss += (loss.data - valid_loss) / (batch_idx + 1)
                pred = output.data.max(1, keepdim=True)[1]
                # compare predictions to true label
                correct += np.sum(np.squeeze(pred.eq(target.data.view_as(pred))).cpu().numpy())
                total += data.size(0)
            
        # print training/validation statistics 
        print('Epoch: {} Time: {} Training Loss: {:.3f} Validation Loss: {:.3f} Val accuracy {:.3f} Correct {:.0f} Total {:.0f}'.format(
            epoch,
            time.time() - start,
            train_loss,
            valid_loss,
            correct / total,
            correct,
            total
            ))
        
        ## TODO: save the model if validation loss has decreased
        if valid_loss < valid_loss_min:
            print('Validation loss decreased ({:.6f} --> {:.6f}).  Saving model ...'.format(
                valid_loss_min,
                valid_loss))
            valid_loss_min = valid_loss
            torch.save(model.state_dict(), save_path)
    # return trained model
    return model
In [56]:
# train the model
from workspace_utils import active_session
from PIL import ImageFile

ImageFile.LOAD_TRUNCATED_IMAGES = True

with active_session():
    model_scratch = train(100, loaders_scratch, model_scratch, optimizer_scratch, 
                          criterion_scratch, use_cuda, 'model_scratch.pt')
Epoch: 1 Time: 113.42174029350281 Training Loss: 3.234 Validation Loss: 2.701 Val accuracy 0.307 Correct 256 Total 835
Validation loss decreased (inf --> 2.700785).  Saving model ...
Epoch: 2 Time: 113.03319835662842 Training Loss: 3.208 Validation Loss: 2.571 Val accuracy 0.339 Correct 283 Total 835
Validation loss decreased (2.700785 --> 2.570942).  Saving model ...
Epoch: 3 Time: 113.07424807548523 Training Loss: 3.225 Validation Loss: 2.584 Val accuracy 0.323 Correct 270 Total 835
Epoch: 4 Time: 112.06094479560852 Training Loss: 3.136 Validation Loss: 2.638 Val accuracy 0.334 Correct 279 Total 835
Epoch: 5 Time: 111.72976636886597 Training Loss: 3.146 Validation Loss: 2.697 Val accuracy 0.322 Correct 269 Total 835
Epoch: 6 Time: 111.6922378540039 Training Loss: 3.087 Validation Loss: 2.540 Val accuracy 0.328 Correct 274 Total 835
Validation loss decreased (2.570942 --> 2.540286).  Saving model ...
Epoch: 7 Time: 113.59004402160645 Training Loss: 3.090 Validation Loss: 2.618 Val accuracy 0.317 Correct 265 Total 835
Epoch: 8 Time: 112.08793592453003 Training Loss: 3.092 Validation Loss: 2.475 Val accuracy 0.345 Correct 288 Total 835
Validation loss decreased (2.540286 --> 2.475280).  Saving model ...
Epoch: 9 Time: 111.77136945724487 Training Loss: 3.055 Validation Loss: 2.520 Val accuracy 0.335 Correct 280 Total 835
Epoch: 10 Time: 111.01392102241516 Training Loss: 3.050 Validation Loss: 2.530 Val accuracy 0.351 Correct 293 Total 835
Epoch: 11 Time: 111.18841433525085 Training Loss: 3.117 Validation Loss: 2.478 Val accuracy 0.358 Correct 299 Total 835
Epoch: 12 Time: 111.17198014259338 Training Loss: 3.051 Validation Loss: 2.568 Val accuracy 0.364 Correct 304 Total 835
Epoch: 13 Time: 111.38805794715881 Training Loss: 3.042 Validation Loss: 2.475 Val accuracy 0.366 Correct 306 Total 835
Validation loss decreased (2.475280 --> 2.474828).  Saving model ...
Epoch: 14 Time: 110.9574978351593 Training Loss: 2.992 Validation Loss: 2.701 Val accuracy 0.362 Correct 302 Total 835
Epoch: 15 Time: 111.48825216293335 Training Loss: 2.979 Validation Loss: 2.381 Val accuracy 0.396 Correct 331 Total 835
Validation loss decreased (2.474828 --> 2.380783).  Saving model ...
Epoch: 16 Time: 112.61486840248108 Training Loss: 2.987 Validation Loss: 2.422 Val accuracy 0.368 Correct 307 Total 835
Epoch: 17 Time: 112.64620923995972 Training Loss: 2.966 Validation Loss: 2.285 Val accuracy 0.390 Correct 326 Total 835
Validation loss decreased (2.380783 --> 2.284956).  Saving model ...
Epoch: 18 Time: 112.9011480808258 Training Loss: 2.972 Validation Loss: 2.330 Val accuracy 0.401 Correct 335 Total 835
Epoch: 19 Time: 112.0682270526886 Training Loss: 2.961 Validation Loss: 2.642 Val accuracy 0.375 Correct 313 Total 835
Epoch: 20 Time: 112.16505646705627 Training Loss: 2.942 Validation Loss: 2.377 Val accuracy 0.380 Correct 317 Total 835
Epoch: 21 Time: 111.90705943107605 Training Loss: 2.930 Validation Loss: 2.265 Val accuracy 0.410 Correct 342 Total 835
Validation loss decreased (2.284956 --> 2.264965).  Saving model ...
Epoch: 22 Time: 113.41980504989624 Training Loss: 2.917 Validation Loss: 2.519 Val accuracy 0.389 Correct 325 Total 835
Epoch: 23 Time: 113.17784786224365 Training Loss: 2.886 Validation Loss: 2.391 Val accuracy 0.363 Correct 303 Total 835
Epoch: 24 Time: 112.63418555259705 Training Loss: 2.916 Validation Loss: 2.269 Val accuracy 0.407 Correct 340 Total 835
Epoch: 25 Time: 111.20785331726074 Training Loss: 2.894 Validation Loss: 2.416 Val accuracy 0.364 Correct 304 Total 835
Epoch: 26 Time: 111.55264973640442 Training Loss: 2.880 Validation Loss: 2.210 Val accuracy 0.414 Correct 346 Total 835
Validation loss decreased (2.264965 --> 2.209511).  Saving model ...
Epoch: 27 Time: 112.68843078613281 Training Loss: 2.860 Validation Loss: 2.284 Val accuracy 0.387 Correct 323 Total 835
Epoch: 28 Time: 113.40783381462097 Training Loss: 2.863 Validation Loss: 2.221 Val accuracy 0.420 Correct 351 Total 835
Epoch: 29 Time: 112.73877096176147 Training Loss: 2.841 Validation Loss: 2.767 Val accuracy 0.398 Correct 332 Total 835
Epoch: 30 Time: 113.70926356315613 Training Loss: 2.827 Validation Loss: 2.215 Val accuracy 0.402 Correct 336 Total 835
Epoch: 31 Time: 113.24742722511292 Training Loss: 2.818 Validation Loss: 2.150 Val accuracy 0.411 Correct 343 Total 835
Validation loss decreased (2.209511 --> 2.150073).  Saving model ...
Epoch: 32 Time: 113.39786386489868 Training Loss: 2.786 Validation Loss: 2.341 Val accuracy 0.426 Correct 356 Total 835
Epoch: 33 Time: 112.21711874008179 Training Loss: 2.814 Validation Loss: 2.246 Val accuracy 0.426 Correct 356 Total 835
Epoch: 34 Time: 111.60044455528259 Training Loss: 2.806 Validation Loss: 2.205 Val accuracy 0.429 Correct 358 Total 835
Epoch: 35 Time: 111.32470440864563 Training Loss: 2.753 Validation Loss: 2.276 Val accuracy 0.434 Correct 362 Total 835
Epoch: 36 Time: 111.1699571609497 Training Loss: 2.786 Validation Loss: 3.202 Val accuracy 0.423 Correct 353 Total 835
Epoch: 37 Time: 110.75214338302612 Training Loss: 2.761 Validation Loss: 2.135 Val accuracy 0.455 Correct 380 Total 835
Validation loss decreased (2.150073 --> 2.135367).  Saving model ...
Epoch: 38 Time: 110.49186873435974 Training Loss: 2.757 Validation Loss: 2.355 Val accuracy 0.437 Correct 365 Total 835
Epoch: 39 Time: 111.23060917854309 Training Loss: 2.793 Validation Loss: 2.030 Val accuracy 0.455 Correct 380 Total 835
Validation loss decreased (2.135367 --> 2.030219).  Saving model ...
Epoch: 40 Time: 110.86963605880737 Training Loss: 2.720 Validation Loss: 2.061 Val accuracy 0.447 Correct 373 Total 835
Epoch: 41 Time: 110.71056199073792 Training Loss: 2.749 Validation Loss: 2.160 Val accuracy 0.457 Correct 382 Total 835
Epoch: 42 Time: 110.59308958053589 Training Loss: 2.724 Validation Loss: 2.043 Val accuracy 0.453 Correct 378 Total 835
Epoch: 43 Time: 112.00689268112183 Training Loss: 2.732 Validation Loss: 2.942 Val accuracy 0.438 Correct 366 Total 835
Epoch: 44 Time: 111.12008953094482 Training Loss: 2.742 Validation Loss: 2.058 Val accuracy 0.460 Correct 384 Total 835
Epoch: 45 Time: 111.19032788276672 Training Loss: 2.708 Validation Loss: 2.455 Val accuracy 0.444 Correct 371 Total 835
Epoch: 46 Time: 112.49147868156433 Training Loss: 2.679 Validation Loss: 2.526 Val accuracy 0.451 Correct 377 Total 835
Epoch: 47 Time: 112.41115593910217 Training Loss: 2.690 Validation Loss: 2.381 Val accuracy 0.474 Correct 396 Total 835
Epoch: 48 Time: 113.22472333908081 Training Loss: 2.721 Validation Loss: 2.073 Val accuracy 0.473 Correct 395 Total 835
Epoch: 49 Time: 112.15785551071167 Training Loss: 2.717 Validation Loss: 2.477 Val accuracy 0.471 Correct 393 Total 835
Epoch: 50 Time: 113.1781792640686 Training Loss: 2.668 Validation Loss: 11.207 Val accuracy 0.446 Correct 372 Total 835
Epoch: 51 Time: 111.66141772270203 Training Loss: 2.668 Validation Loss: 2.108 Val accuracy 0.467 Correct 390 Total 835
Epoch: 52 Time: 111.7741219997406 Training Loss: 2.669 Validation Loss: 2.176 Val accuracy 0.472 Correct 394 Total 835
Epoch: 53 Time: 110.88514852523804 Training Loss: 2.674 Validation Loss: 2.039 Val accuracy 0.448 Correct 374 Total 835
Epoch: 54 Time: 111.6361632347107 Training Loss: 2.628 Validation Loss: 2.253 Val accuracy 0.475 Correct 397 Total 835
Epoch: 55 Time: 111.27090239524841 Training Loss: 2.640 Validation Loss: 2.906 Val accuracy 0.491 Correct 410 Total 835
Epoch: 56 Time: 111.48240399360657 Training Loss: 2.629 Validation Loss: 2.047 Val accuracy 0.480 Correct 401 Total 835
Epoch: 57 Time: 111.07416677474976 Training Loss: 2.580 Validation Loss: 2.017 Val accuracy 0.501 Correct 418 Total 835
Validation loss decreased (2.030219 --> 2.016749).  Saving model ...
Epoch: 58 Time: 111.32993936538696 Training Loss: 2.633 Validation Loss: 2.017 Val accuracy 0.498 Correct 416 Total 835
Epoch: 59 Time: 111.36694717407227 Training Loss: 2.594 Validation Loss: 1.918 Val accuracy 0.515 Correct 430 Total 835
Validation loss decreased (2.016749 --> 1.917668).  Saving model ...
Epoch: 60 Time: 111.39702439308167 Training Loss: 2.580 Validation Loss: 2.390 Val accuracy 0.475 Correct 397 Total 835
Epoch: 61 Time: 111.56429243087769 Training Loss: 2.630 Validation Loss: 2.059 Val accuracy 0.467 Correct 390 Total 835
Epoch: 62 Time: 110.80414581298828 Training Loss: 2.592 Validation Loss: 1.983 Val accuracy 0.479 Correct 400 Total 835
Epoch: 63 Time: 111.16559672355652 Training Loss: 2.593 Validation Loss: 1.923 Val accuracy 0.501 Correct 418 Total 835
Epoch: 64 Time: 111.13597512245178 Training Loss: 2.589 Validation Loss: 2.046 Val accuracy 0.492 Correct 411 Total 835
Epoch: 65 Time: 111.6567895412445 Training Loss: 2.588 Validation Loss: 2.004 Val accuracy 0.487 Correct 407 Total 835
Epoch: 66 Time: 112.04625129699707 Training Loss: 2.544 Validation Loss: 2.044 Val accuracy 0.469 Correct 392 Total 835
Epoch: 67 Time: 112.61690521240234 Training Loss: 2.547 Validation Loss: 2.550 Val accuracy 0.501 Correct 418 Total 835
Epoch: 68 Time: 112.11928129196167 Training Loss: 2.571 Validation Loss: 2.004 Val accuracy 0.491 Correct 410 Total 835
Epoch: 69 Time: 111.64204049110413 Training Loss: 2.521 Validation Loss: 1.924 Val accuracy 0.483 Correct 403 Total 835
Epoch: 70 Time: 111.55069732666016 Training Loss: 2.531 Validation Loss: 1.946 Val accuracy 0.489 Correct 408 Total 835
Epoch: 71 Time: 111.81630063056946 Training Loss: 2.566 Validation Loss: 1.939 Val accuracy 0.485 Correct 405 Total 835
Epoch: 72 Time: 112.51158094406128 Training Loss: 2.553 Validation Loss: 2.083 Val accuracy 0.504 Correct 421 Total 835
Epoch: 73 Time: 112.90292978286743 Training Loss: 2.483 Validation Loss: 2.366 Val accuracy 0.453 Correct 378 Total 835
Epoch: 74 Time: 113.0706148147583 Training Loss: 2.541 Validation Loss: 1.927 Val accuracy 0.503 Correct 420 Total 835
Epoch: 75 Time: 112.04956555366516 Training Loss: 2.524 Validation Loss: 3.374 Val accuracy 0.499 Correct 417 Total 835
Epoch: 76 Time: 113.13687705993652 Training Loss: 2.515 Validation Loss: 1.941 Val accuracy 0.509 Correct 425 Total 835
Epoch: 77 Time: 113.35320830345154 Training Loss: 2.533 Validation Loss: 2.998 Val accuracy 0.521 Correct 435 Total 835
Epoch: 78 Time: 113.04463744163513 Training Loss: 2.485 Validation Loss: 2.976 Val accuracy 0.491 Correct 410 Total 835
Epoch: 79 Time: 111.62535119056702 Training Loss: 2.499 Validation Loss: 3.182 Val accuracy 0.486 Correct 406 Total 835
Epoch: 80 Time: 112.88961982727051 Training Loss: 2.477 Validation Loss: 3.770 Val accuracy 0.503 Correct 420 Total 835
Epoch: 81 Time: 111.93040013313293 Training Loss: 2.465 Validation Loss: 3.788 Val accuracy 0.497 Correct 415 Total 835
Epoch: 82 Time: 112.38668918609619 Training Loss: 2.511 Validation Loss: 2.009 Val accuracy 0.492 Correct 411 Total 835
Epoch: 83 Time: 112.59733128547668 Training Loss: 2.494 Validation Loss: 2.562 Val accuracy 0.478 Correct 399 Total 835
Epoch: 84 Time: 111.69334530830383 Training Loss: 2.507 Validation Loss: 4.924 Val accuracy 0.510 Correct 426 Total 835
Epoch: 85 Time: 112.90107369422913 Training Loss: 2.511 Validation Loss: 1.957 Val accuracy 0.525 Correct 438 Total 835
Epoch: 86 Time: 112.97108507156372 Training Loss: 2.478 Validation Loss: 2.073 Val accuracy 0.474 Correct 396 Total 835
Epoch: 87 Time: 113.3069794178009 Training Loss: 2.507 Validation Loss: 1.879 Val accuracy 0.497 Correct 415 Total 835
Validation loss decreased (1.917668 --> 1.879154).  Saving model ...
Epoch: 88 Time: 112.98117923736572 Training Loss: 2.473 Validation Loss: 2.242 Val accuracy 0.495 Correct 413 Total 835
Epoch: 89 Time: 112.93390154838562 Training Loss: 2.480 Validation Loss: 2.286 Val accuracy 0.489 Correct 408 Total 835
Epoch: 90 Time: 113.38753604888916 Training Loss: 2.476 Validation Loss: 2.039 Val accuracy 0.534 Correct 446 Total 835
Epoch: 91 Time: 113.0846176147461 Training Loss: 2.456 Validation Loss: 2.418 Val accuracy 0.520 Correct 434 Total 835
Epoch: 92 Time: 113.06859540939331 Training Loss: 2.445 Validation Loss: 1.891 Val accuracy 0.521 Correct 435 Total 835
Epoch: 93 Time: 113.01021122932434 Training Loss: 2.467 Validation Loss: 2.504 Val accuracy 0.508 Correct 424 Total 835
Epoch: 94 Time: 112.53139519691467 Training Loss: 2.469 Validation Loss: 2.201 Val accuracy 0.496 Correct 414 Total 835
Epoch: 95 Time: 113.46156406402588 Training Loss: 2.465 Validation Loss: 1.841 Val accuracy 0.535 Correct 447 Total 835
Validation loss decreased (1.879154 --> 1.840984).  Saving model ...
Epoch: 96 Time: 112.65268468856812 Training Loss: 2.461 Validation Loss: 2.267 Val accuracy 0.525 Correct 438 Total 835
Epoch: 97 Time: 113.59708070755005 Training Loss: 2.435 Validation Loss: 2.585 Val accuracy 0.526 Correct 439 Total 835
Epoch: 98 Time: 112.53133726119995 Training Loss: 2.456 Validation Loss: 2.043 Val accuracy 0.523 Correct 437 Total 835
Epoch: 99 Time: 112.68015313148499 Training Loss: 2.452 Validation Loss: 1.947 Val accuracy 0.519 Correct 433 Total 835
Epoch: 100 Time: 112.80211043357849 Training Loss: 2.457 Validation Loss: 1.936 Val accuracy 0.499 Correct 417 Total 835
In [13]:
# load the model that got the best validation accuracy
model_scratch.load_state_dict(torch.load('model_scratch.pt'))

(IMPLEMENTATION) Test the Model

Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 10%.

In [14]:
def test(loaders, model, criterion, use_cuda):

    # monitor test loss and accuracy
    test_loss = 0.
    correct = 0.
    total = 0.

    model.eval()
    for batch_idx, (data, target) in enumerate(loaders['test']):
        # move to GPU
        if use_cuda:
            data, target = data.cuda(), target.cuda()
        # forward pass: compute predicted outputs by passing inputs to the model
        output = model(data)
        # calculate the loss
        loss = criterion(output, target)
        # update average test loss 
        test_loss = test_loss + ((1 / (batch_idx + 1)) * (loss.data - test_loss))
        # convert output probabilities to predicted class
        pred = output.data.max(1, keepdim=True)[1]
        # compare predictions to true label
        correct += np.sum(np.squeeze(pred.eq(target.data.view_as(pred))).cpu().numpy())
        total += data.size(0)
            
    print('Test Loss: {:.6f}\n'.format(test_loss))

    print('\nTest Accuracy: %2d%% (%2d/%2d)' % (
        100. * correct / total, correct, total))
In [17]:
# call test function    
test(loaders_scratch, model_scratch, criterion_scratch, use_cuda)
Test Loss: 1.914108


Test Accuracy: 53% (449/836)

Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)

You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.

(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset

Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dogImages/train, dogImages/valid, and dogImages/test, respectively).

If you like, you are welcome to use the same data loaders from the previous step, when you created a CNN from scratch.

In [15]:
base_dir = '/data/dog_images'

train_transform = transforms.Compose([transforms.RandomResizedCrop(224),
                                      transforms.RandomRotation(30),
                                      transforms.RandomHorizontalFlip(),
                                      transforms.ToTensor(),
                                      transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                         std=[0.229, 0.224, 0.225])
                                     ])

test_transform = transforms.Compose([transforms.Resize(256),
                                     transforms.CenterCrop(224),
                                     transforms.ToTensor(),
                                     transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                         std=[0.229, 0.224, 0.225])
                                     ])

train_data = datasets.ImageFolder(base_dir + '/train', transform=train_transform)
valiadation_data = datasets.ImageFolder(base_dir + '/valid', transform=test_transform)
test_data = datasets.ImageFolder(base_dir + '/test', transform=test_transform)

train_loader = torch.utils.data.DataLoader(train_data, batch_size=10, shuffle=True)
validation_loader = torch.utils.data.DataLoader(valiadation_data, batch_size=10, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=10, shuffle=True)

loaders_transfer = {
    'train': train_loader,
    'valid': validation_loader,
    'test': test_loader
}

(IMPLEMENTATION) Model Architecture

Use transfer learning to create a CNN to classify dog breed. Use the code cell below, and save your initialized model as the variable model_transfer.

In [16]:
import torchvision.models as models
import torch.nn as nn

model_transfer = models.vgg16(pretrained=True)
for param in model_transfer.parameters():
    param.requires_grad = False

model_transfer.classifier[6] = nn.Linear(model_transfer.classifier[6].in_features, 133)

if use_cuda:
    model_transfer = model_transfer.cuda()

Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.

Answer:

From what we found earlier it looks like the VGG16 did a pretty good job at classifying dog breeds already. So, I began by following what we did in the transfer learning solution for flowers, but decided I only need to retrain and replace the last layer. Bascially, it's already good at detecting dog breeds, but the values that it outputs have changed so it shouldn't need a lot of manipulation.

(IMPLEMENTATION) Specify Loss Function and Optimizer

Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_transfer, and the optimizer as optimizer_transfer below.

In [17]:
import torch.optim as optim

criterion_transfer = nn.CrossEntropyLoss()
optimizer_transfer = optim.Adam(model_transfer.classifier[6].
        parameters(), lr = 0.0001)

(IMPLEMENTATION) Train and Validate the Model

Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_transfer.pt'.

In [36]:
# train the model
# train the model
from workspace_utils import active_session
from PIL import ImageFile

ImageFile.LOAD_TRUNCATED_IMAGES = True

with active_session():
    model_transfer = train(100, loaders_transfer, model_transfer, optimizer_transfer, criterion_transfer, use_cuda, 'model_transfer.pt')
---------------------------------------------------------------------------
KeyboardInterrupt                         Traceback (most recent call last)
<ipython-input-36-7338d34a3133> in <module>()
      7 
      8 with active_session():
----> 9     model_transfer = train(100, loaders_transfer, model_transfer, optimizer_transfer, criterion_transfer, use_cuda, 'model_transfer.pt')

<ipython-input-35-1896e00c67a9> in train(n_epochs, loaders, model, optimizer, criterion, use_cuda, save_path)
     19             # move to GPU
     20             if use_cuda:
---> 21                 data, target = data.cuda(), target.cuda()
     22             ## find the loss and update the model parameters accordingly
     23             ## record the average training loss, using something like

KeyboardInterrupt: 
In [18]:
# load the model that got the best validation accuracy (uncomment the line below)
model_transfer.load_state_dict(torch.load('model_transfer.pt'))

(IMPLEMENTATION) Test the Model

Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 60%.

In [38]:
test(loaders_transfer, model_transfer, criterion_transfer, use_cuda)
Test Loss: 0.605507


Test Accuracy: 83% (700/836)

(IMPLEMENTATION) Predict Dog Breed with the Model

Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan hound, etc) that is predicted by your model.

In [19]:
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.

# list of class names by index, i.e. a name can be accessed like class_names[0]
class_names = [item[4:].replace("_", " ") for item in train_data.classes]

def predict_breed_transfer(img_path):
    return class_names[classify(model_transfer, img_path)]

Step 5: Write your Algorithm

Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,

  • if a dog is detected in the image, return the predicted breed.
  • if a human is detected in the image, return the resembling dog breed.
  • if neither is detected in the image, provide output that indicates an error.

You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and human_detector functions developed above. You are required to use your CNN from Step 4 to predict dog breed.

Some sample output for our algorithm is provided below, but feel free to design your own user experience!

Sample Human Output

(IMPLEMENTATION) Write your Algorithm

In [22]:
### TODO: Write your algorithm.
### Feel free to use as many code cells as needed.
from enum import Enum

class Detected_Entity(Enum):
    BAMBOOZLE = 0
    HOOMAN = 1
    DOGGO = 2

def detect_entity(img_path):
    if dog_detector(img_path):
        return Detected_Entity.DOGGO
    elif face_detector(img_path):
        return Detected_Entity.HOOMAN
    else:
        return Detected_Entity.BAMBOOZLE

greetings = {
    Detected_Entity.BAMBOOZLE: 'Hello fren,',
    Detected_Entity.DOGGO: 'Hello doggo fren,',
    Detected_Entity.HOOMAN: 'Hello hooman fren,'
}

intro = {
    Detected_Entity.BAMBOOZLE: 'this do me a heckin bamboozle, is it a snek?',
    Detected_Entity.DOGGO: 'I think you is a ...',
    Detected_Entity.HOOMAN: 'I think you look like a ...'
}

def show_image(img_path):
    img = cv2.imread(img_path)
    cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    plt.imshow(cv_rgb)
    plt.show()

def run_app(img_path):
    entity = detect_entity(img_path)
    print(greetings[entity])
    show_image(img_path)
    print(intro[entity])
    if entity != Detected_Entity.BAMBOOZLE:
        print(predict_breed_transfer(img_path))
    print('')
    print('')
    print('')

Step 6: Test Your Algorithm

In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?

(IMPLEMENTATION) Test Your Algorithm on Sample Images!

Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.

Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.

Answer: (Three possible points for improvement)

In [27]:
## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.

## suggested code, below
from os import listdir
from os.path import isfile, join

folder = 'images/random'
for file in [f for f in listdir(folder) if isfile(join(folder, f))]:
    run_app(folder + '/' + file)

#for file in np.hstack((human_files[:3], dog_files[:3])):
#    run_app(file)
Hello doggo fren,
I think you is a ...
Dachshund



Hello doggo fren,
I think you is a ...
Golden retriever



Hello hooman fren,
I think you look like a ...
Silky terrier



Hello doggo fren,
I think you is a ...
Greater swiss mountain dog



Hello hooman fren,
I think you look like a ...
Pharaoh hound



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello hooman fren,
I think you look like a ...
Wirehaired pointing griffon



Hello doggo fren,
I think you is a ...
American foxhound



Hello doggo fren,
I think you is a ...
Italian greyhound



Hello doggo fren,
I think you is a ...
Cardigan welsh corgi



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello doggo fren,
I think you is a ...
Manchester terrier



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello hooman fren,
I think you look like a ...
Dogue de bordeaux



Hello doggo fren,
I think you is a ...
Belgian sheepdog



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello doggo fren,
I think you is a ...
Neapolitan mastiff



Hello hooman fren,
I think you look like a ...
Chinese crested



Hello doggo fren,
I think you is a ...
Alaskan malamute



Hello hooman fren,
I think you look like a ...
Silky terrier



Hello hooman fren,
I think you look like a ...
Dachshund



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello hooman fren,
I think you look like a ...
Dogue de bordeaux



Hello hooman fren,
I think you look like a ...
Chinese crested



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello fren,
this do me a heckin bamboozle, is it a snek?



Hello hooman fren,
I think you look like a ...
Silky terrier



Hello doggo fren,
I think you is a ...
Dachshund